This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
MedShift: Implicit Conditional Transport for X-Ray Domain Adaptation
Created by
Haebom
Author
Francisco Caetano, Christiaan Viviers, Peter HH de With, Fons van der Sommen
Outline
Synthetic medical data offers a scalable solution for training robust models, but significant domain gaps limit generalization to real-world clinical settings. This paper addresses the problem of cross-domain translation between synthetic and real-world head X images by focusing on resolving discrepancies in attenuation behavior, noise characteristics, and soft tissue representation. We propose MedShift, a unified class-conditional generative model based on Flow Matching and Schrodinger Bridges. This model enables high-fidelity, unpaired image translation across multiple domains. Unlike previous approaches that require domain-specific training or rely on paired data, MedShift learns a shared, domain-independent latent space and enables seamless translation between all observed domain pairs during training. We also benchmark domain translation models by introducing X-DigiSkull, a novel dataset consisting of aligned synthetic and real skull X lines from various radiation doses. Experimental results demonstrate that MedShift delivers robust performance despite a smaller model size compared to diffusion-based approaches, and that it can be tuned to prioritize either perceptual fidelity or structural consistency during inference, making it a scalable and generalizable solution for domain adaptation in medical imaging. The code and dataset are available at https://caetas.github.io/medshift.html .
Takeaways, Limitations
•
Takeaways:
◦
High-fidelity, unpaired image transformation between synthetic and real medical images is possible through our proposed MedShift, a unified class conditional generative model based on Flow Matching and Schrodinger Bridges.
◦
Shared domain-independent latent space learning without domain-specific training or dependence on paired data.
◦
Flexibility is achieved by allowing the prioritization of either perceptual fidelity or structural consistency during inference.
◦
Provides robust performance with smaller model sizes than diffusion-based approaches.
◦
New dataset X-DigiSkull enables benchmarking of domain transformation models.
◦
Providing scalable and generalizable domain-adapted solutions
•
Limitations:
◦
Limitations is not explicitly mentioned in the paper. Further research is needed to further verify its generalization performance in real-world clinical settings and its applicability to various medical imaging modalities.
